drivers/ntb/ntb_transport.c

   1 /*
   2  * This file is provided under a dual BSD/GPLv2 license.  When using or
   3  *   redistributing this file, you may do so under either license.
   4  *
   5  *   GPL LICENSE SUMMARY
   6  *
   7  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
   8  *
   9  *   This program is free software; you can redistribute it and/or modify
  10  *   it under the terms of version 2 of the GNU General Public License as
  11  *   published by the Free Software Foundation.
  12  *
  13  *   BSD LICENSE
  14  *
  15  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
  16  *
  17  *   Redistribution and use in source and binary forms, with or without
  18  *   modification, are permitted provided that the following conditions
  19  *   are met:
  20  *
  21  *     * Redistributions of source code must retain the above copyright
  22  *       notice, this list of conditions and the following disclaimer.
  23  *     * Redistributions in binary form must reproduce the above copy
  24  *       notice, this list of conditions and the following disclaimer in
  25  *       the documentation and/or other materials provided with the
  26  *       distribution.
  27  *     * Neither the name of Intel Corporation nor the names of its
  28  *       contributors may be used to endorse or promote products derived
  29  *       from this software without specific prior written permission.
  30  *
  31  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  32  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  33  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  34  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  35  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  36  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  37  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  38  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  39  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  40  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  41  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  42  *
  43  * Intel PCIe NTB Linux driver
  44  *
  45  * Contact Information:
  46  * Jon Mason <jon.mason@intel.com>
  47  */
  48 #include <linux/debugfs.h>
  49 #include <linux/delay.h>
  50 #include <linux/dmaengine.h>
  51 #include <linux/dma-mapping.h>
  52 #include <linux/errno.h>
  53 #include <linux/export.h>
  54 #include <linux/interrupt.h>
  55 #include <linux/module.h>
  56 #include <linux/pci.h>
  57 #include <linux/slab.h>
  58 #include <linux/types.h>
  59 #include "ntb_hw.h"
  60
  61 #define NTB_TRANSPORT_VERSION   3
  62
  63 static unsigned int transport_mtu = 0x401E;
  64 module_param(transport_mtu, uint, 0644);
  65 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  66
  67 static unsigned char max_num_clients;
  68 module_param(max_num_clients, byte, 0644);
  69 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  70
  71 static unsigned int copy_bytes = 1024;
  72 module_param(copy_bytes, uint, 0644);
  73 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  74
  75 struct ntb_queue_entry {
  76         /* ntb_queue list reference */
  77         struct list_head entry;
  78         /* pointers to data to be transfered */
  79         void *cb_data;
  80         void *buf;
  81         unsigned int len;
  82         unsigned int flags;
  83
  84         struct ntb_transport_qp *qp;
  85         union {
  86                 struct ntb_payload_header __iomem *tx_hdr;
  87                 struct ntb_payload_header *rx_hdr;
  88         };
  89         unsigned int index;
  90 };
  91
  92 struct ntb_rx_info {
  93         unsigned int entry;
  94 };
  95
  96 struct ntb_transport_qp {
  97         struct ntb_transport *transport;
  98         struct ntb_device *ndev;
  99         void *cb_data;
 100         struct dma_chan *dma_chan;
 101
 102         bool client_ready;
 103         bool qp_link;
 104         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 105
 106         struct ntb_rx_info __iomem *rx_info;
 107         struct ntb_rx_info *remote_rx_info;
 108
 109         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 110                            void *data, int len);
 111         struct list_head tx_free_q;
 112         spinlock_t ntb_tx_free_q_lock;
 113         void __iomem *tx_mw;
 114         dma_addr_t tx_mw_phys;
 115         unsigned int tx_index;
 116         unsigned int tx_max_entry;
 117         unsigned int tx_max_frame;
 118
 119         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 120                            void *data, int len);
 121         struct list_head rx_pend_q;
 122         struct list_head rx_free_q;
 123         spinlock_t ntb_rx_pend_q_lock;
 124         spinlock_t ntb_rx_free_q_lock;
 125         void *rx_buff;
 126         unsigned int rx_index;
 127         unsigned int rx_max_entry;
 128         unsigned int rx_max_frame;
 129         dma_cookie_t last_cookie;
 130
 131         void (*event_handler)(void *data, int status);
 132         struct delayed_work link_work;
 133         struct work_struct link_cleanup;
 134
 135         struct dentry *debugfs_dir;
 136         struct dentry *debugfs_stats;
 137
 138         /* Stats */
 139         u64 rx_bytes;
 140         u64 rx_pkts;
 141         u64 rx_ring_empty;
 142         u64 rx_err_no_buf;
 143         u64 rx_err_oflow;
 144         u64 rx_err_ver;
 145         u64 rx_memcpy;
 146         u64 rx_async;
 147         u64 tx_bytes;
 148         u64 tx_pkts;
 149         u64 tx_ring_full;
 150         u64 tx_err_no_buf;
 151         u64 tx_memcpy;
 152         u64 tx_async;
 153 };
 154
 155 struct ntb_transport_mw {
 156         size_t size;
 157         void *virt_addr;
 158         dma_addr_t dma_addr;
 159 };
 160
 161 struct ntb_transport_client_dev {
 162         struct list_head entry;
 163         struct device dev;
 164 };
 165
 166 struct ntb_transport {
 167         struct list_head entry;
 168         struct list_head client_devs;
 169
 170         struct ntb_device *ndev;
 171         struct ntb_transport_mw *mw;
 172         struct ntb_transport_qp *qps;
 173         unsigned int max_qps;
 174         unsigned long qp_bitmap;
 175         bool transport_link;
 176         struct delayed_work link_work;
 177         struct work_struct link_cleanup;
 178 };
 179
 180 enum {
 181         DESC_DONE_FLAG = 1 << 0,
 182         LINK_DOWN_FLAG = 1 << 1,
 183 };
 184
 185 struct ntb_payload_header {
 186         unsigned int ver;
 187         unsigned int len;
 188         unsigned int flags;
 189 };
 190
 191 enum {
 192         VERSION = 0,
 193         QP_LINKS,
 194         NUM_QPS,
 195         NUM_MWS,
 196         MW0_SZ_HIGH,
 197         MW0_SZ_LOW,
 198         MW1_SZ_HIGH,
 199         MW1_SZ_LOW,
 200         MAX_SPAD,
 201 };
 202
 203 #define QP_TO_MW(ndev, qp)      ((qp) % ntb_max_mw(ndev))
 204 #define NTB_QP_DEF_NUM_ENTRIES  100
 205 #define NTB_LINK_DOWN_TIMEOUT   10
 206
 207 static int ntb_match_bus(struct device *dev, struct device_driver *drv)
 208 {
 209         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 210 }
 211
 212 static int ntb_client_probe(struct device *dev)
 213 {
 214         const struct ntb_client *drv = container_of(dev->driver,
 215                                                     struct ntb_client, driver);
 216         struct pci_dev *pdev = container_of(dev->parent, struct pci_dev, dev);
 217         int rc = -EINVAL;
 218
 219         get_device(dev);
 220         if (drv && drv->probe)
 221                 rc = drv->probe(pdev);
 222         if (rc)
 223                 put_device(dev);
 224
 225         return rc;
 226 }
 227
 228 static int ntb_client_remove(struct device *dev)
 229 {
 230         const struct ntb_client *drv = container_of(dev->driver,
 231                                                     struct ntb_client, driver);
 232         struct pci_dev *pdev = container_of(dev->parent, struct pci_dev, dev);
 233
 234         if (drv && drv->remove)
 235                 drv->remove(pdev);
 236
 237         put_device(dev);
 238
 239         return 0;
 240 }
 241
 242 static struct bus_type ntb_bus_type = {
 243         .name = "ntb_bus",
 244         .match = ntb_match_bus,
 245         .probe = ntb_client_probe,
 246         .remove = ntb_client_remove,
 247 };
 248
 249 static LIST_HEAD(ntb_transport_list);
 250
 251 static int ntb_bus_init(struct ntb_transport *nt)
 252 {
 253         if (list_empty(&ntb_transport_list)) {
 254                 int rc = bus_register(&ntb_bus_type);
 255                 if (rc)
 256                         return rc;
 257         }
 258
 259         list_add(&nt->entry, &ntb_transport_list);
 260
 261         return 0;
 262 }
 263
 264 static void ntb_bus_remove(struct ntb_transport *nt)
 265 {
 266         struct ntb_transport_client_dev *client_dev, *cd;
 267
 268         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 269                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 270                         dev_name(&client_dev->dev));
 271                 list_del(&client_dev->entry);
 272                 device_unregister(&client_dev->dev);
 273         }
 274
 275         list_del(&nt->entry);
 276
 277         if (list_empty(&ntb_transport_list))
 278                 bus_unregister(&ntb_bus_type);
 279 }
 280
 281 static void ntb_client_release(struct device *dev)
 282 {
 283         struct ntb_transport_client_dev *client_dev;
 284         client_dev = container_of(dev, struct ntb_transport_client_dev, dev);
 285
 286         kfree(client_dev);
 287 }
 288
 289 /**
 290  * ntb_unregister_client_dev - Unregister NTB client device
 291  * @device_name: Name of NTB client device
 292  *
 293  * Unregister an NTB client device with the NTB transport layer
 294  */
 295 void ntb_unregister_client_dev(char *device_name)
 296 {
 297         struct ntb_transport_client_dev *client, *cd;
 298         struct ntb_transport *nt;
 299
 300         list_for_each_entry(nt, &ntb_transport_list, entry)
 301                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 302                         if (!strncmp(dev_name(&client->dev), device_name,
 303                                      strlen(device_name))) {
 304                                 list_del(&client->entry);
 305                                 device_unregister(&client->dev);
 306                         }
 307 }
 308 EXPORT_SYMBOL_GPL(ntb_unregister_client_dev);
 309
 310 /**
 311  * ntb_register_client_dev - Register NTB client device
 312  * @device_name: Name of NTB client device
 313  *
 314  * Register an NTB client device with the NTB transport layer
 315  */
 316 int ntb_register_client_dev(char *device_name)
 317 {
 318         struct ntb_transport_client_dev *client_dev;
 319         struct ntb_transport *nt;
 320         int rc, i = 0;
 321
 322         if (list_empty(&ntb_transport_list))
 323                 return -ENODEV;
 324
 325         list_for_each_entry(nt, &ntb_transport_list, entry) {
 326                 struct device *dev;
 327
 328                 client_dev = kzalloc(sizeof(struct ntb_transport_client_dev),
 329                                      GFP_KERNEL);
 330                 if (!client_dev) {
 331                         rc = -ENOMEM;
 332                         goto err;
 333                 }
 334
 335                 dev = &client_dev->dev;
 336
 337                 /* setup and register client devices */
 338                 dev_set_name(dev, "%s%d", device_name, i);
 339                 dev->bus = &ntb_bus_type;
 340                 dev->release = ntb_client_release;
 341                 dev->parent = &ntb_query_pdev(nt->ndev)->dev;
 342
 343                 rc = device_register(dev);
 344                 if (rc) {
 345                         kfree(client_dev);
 346                         goto err;
 347                 }
 348
 349                 list_add_tail(&client_dev->entry, &nt->client_devs);
 350                 i++;
 351         }
 352
 353         return 0;
 354
 355 err:
 356         ntb_unregister_client_dev(device_name);
 357
 358         return rc;
 359 }
 360 EXPORT_SYMBOL_GPL(ntb_register_client_dev);
 361
 362 /**
 363  * ntb_register_client - Register NTB client driver
 364  * @drv: NTB client driver to be registered
 365  *
 366  * Register an NTB client driver with the NTB transport layer
 367  *
 368  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 369  */
 370 int ntb_register_client(struct ntb_client *drv)
 371 {
 372         drv->driver.bus = &ntb_bus_type;
 373
 374         if (list_empty(&ntb_transport_list))
 375                 return -ENODEV;
 376
 377         return driver_register(&drv->driver);
 378 }
 379 EXPORT_SYMBOL_GPL(ntb_register_client);
 380
 381 /**
 382  * ntb_unregister_client - Unregister NTB client driver
 383  * @drv: NTB client driver to be unregistered
 384  *
 385  * Unregister an NTB client driver with the NTB transport layer
 386  *
 387  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 388  */
 389 void ntb_unregister_client(struct ntb_client *drv)
 390 {
 391         driver_unregister(&drv->driver);
 392 }
 393 EXPORT_SYMBOL_GPL(ntb_unregister_client);
 394
 395 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 396                             loff_t *offp)
 397 {
 398         struct ntb_transport_qp *qp;
 399         char *buf;
 400         ssize_t ret, out_offset, out_count;
 401
 402         out_count = 1000;
 403
 404         buf = kmalloc(out_count, GFP_KERNEL);
 405         if (!buf)
 406                 return -ENOMEM;
 407
 408         qp = filp->private_data;
 409         out_offset = 0;
 410         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 411                                "NTB QP stats\n");
 412         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 413                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 414         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 415                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 416         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 417                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 418         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 419                                "rx_async - \t%llu\n", qp->rx_async);
 420         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 421                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 422         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 423                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 424         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 425                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 426         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 427                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 428         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 429                                "rx_buff - \t%p\n", qp->rx_buff);
 430         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 431                                "rx_index - \t%u\n", qp->rx_index);
 432         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 433                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 434
 435         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 436                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 437         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 438                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 439         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 440                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 441         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 442                                "tx_async - \t%llu\n", qp->tx_async);
 443         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 444                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 445         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 446                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 447         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 448                                "tx_mw - \t%p\n", qp->tx_mw);
 449         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 450                                "tx_index - \t%u\n", qp->tx_index);
 451         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 452                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 453
 454         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 455                                "\nQP Link %s\n", (qp->qp_link == NTB_LINK_UP) ?
 456                                "Up" : "Down");
 457         if (out_offset > out_count)
 458                 out_offset = out_count;
 459
 460         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 461         kfree(buf);
 462         return ret;
 463 }
 464
 465 static const struct file_operations ntb_qp_debugfs_stats = {
 466         .owner = THIS_MODULE,
 467         .open = simple_open,
 468         .read = debugfs_read,
 469 };
 470
 471 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 472                          struct list_head *list)
 473 {
 474         unsigned long flags;
 475
 476         spin_lock_irqsave(lock, flags);
 477         list_add_tail(entry, list);
 478         spin_unlock_irqrestore(lock, flags);
 479 }
 480
 481 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 482                                            struct list_head *list)
 483 {
 484         struct ntb_queue_entry *entry;
 485         unsigned long flags;
 486
 487         spin_lock_irqsave(lock, flags);
 488         if (list_empty(list)) {
 489                 entry = NULL;
 490                 goto out;
 491         }
 492         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 493         list_del(&entry->entry);
 494 out:
 495         spin_unlock_irqrestore(lock, flags);
 496
 497         return entry;
 498 }
 499
 500 static void ntb_transport_setup_qp_mw(struct ntb_transport *nt,
 501                                       unsigned int qp_num)
 502 {
 503         struct ntb_transport_qp *qp = &nt->qps[qp_num];
 504         unsigned int rx_size, num_qps_mw;
 505         u8 mw_num, mw_max;
 506         unsigned int i;
 507
 508         mw_max = ntb_max_mw(nt->ndev);
 509         mw_num = QP_TO_MW(nt->ndev, qp_num);
 510
 511         WARN_ON(nt->mw[mw_num].virt_addr == NULL);
 512
 513         if (nt->max_qps % mw_max && mw_num < nt->max_qps % mw_max)
 514                 num_qps_mw = nt->max_qps / mw_max + 1;
 515         else
 516                 num_qps_mw = nt->max_qps / mw_max;
 517
 518         rx_size = (unsigned int) nt->mw[mw_num].size / num_qps_mw;
 519         qp->rx_buff = nt->mw[mw_num].virt_addr + qp_num / mw_max * rx_size;
 520         rx_size -= sizeof(struct ntb_rx_info);
 521
 522         qp->remote_rx_info = qp->rx_buff + rx_size;
 523
 524         /* Due to housekeeping, there must be atleast 2 buffs */
 525         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 526         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 527         qp->rx_index = 0;
 528
 529         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 530
 531         /* setup the hdr offsets with 0's */
 532         for (i = 0; i < qp->rx_max_entry; i++) {
 533                 void *offset = qp->rx_buff + qp->rx_max_frame * (i + 1) -
 534                                sizeof(struct ntb_payload_header);
 535                 memset(offset, 0, sizeof(struct ntb_payload_header));
 536         }
 537
 538         qp->rx_pkts = 0;
 539         qp->tx_pkts = 0;
 540         qp->tx_index = 0;
 541 }
 542
 543 static void ntb_free_mw(struct ntb_transport *nt, int num_mw)
 544 {
 545         struct ntb_transport_mw *mw = &nt->mw[num_mw];
 546         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 547
 548         if (!mw->virt_addr)
 549                 return;
 550
 551         dma_free_coherent(&pdev->dev, mw->size, mw->virt_addr, mw->dma_addr);
 552         mw->virt_addr = NULL;
 553 }
 554
 555 static int ntb_set_mw(struct ntb_transport *nt, int num_mw, unsigned int size)
 556 {
 557         struct ntb_transport_mw *mw = &nt->mw[num_mw];
 558         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 559
 560         /* No need to re-setup */
 561         if (mw->size == ALIGN(size, 4096))
 562                 return 0;
 563
 564         if (mw->size != 0)
 565                 ntb_free_mw(nt, num_mw);
 566
 567         /* Alloc memory for receiving data.  Must be 4k aligned */
 568         mw->size = ALIGN(size, 4096);
 569
 570         mw->virt_addr = dma_alloc_coherent(&pdev->dev, mw->size, &mw->dma_addr,
 571                                            GFP_KERNEL);
 572         if (!mw->virt_addr) {
 573                 mw->size = 0;
 574                 dev_err(&pdev->dev, "Unable to allocate MW buffer of size %d\n",
 575                        (int) mw->size);
 576                 return -ENOMEM;
 577         }
 578
 579         /* Notify HW the memory location of the receive buffer */
 580         ntb_set_mw_addr(nt->ndev, num_mw, mw->dma_addr);
 581
 582         return 0;
 583 }
 584
 585 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 586 {
 587         struct ntb_transport *nt = qp->transport;
 588         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 589
 590         if (qp->qp_link == NTB_LINK_DOWN) {
 591                 cancel_delayed_work_sync(&qp->link_work);
 592                 return;
 593         }
 594
 595         if (qp->event_handler)
 596                 qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
 597
 598         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
 599         qp->qp_link = NTB_LINK_DOWN;
 600 }
 601
 602 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 603 {
 604         struct ntb_transport_qp *qp = container_of(work,
 605                                                    struct ntb_transport_qp,
 606                                                    link_cleanup);
 607         struct ntb_transport *nt = qp->transport;
 608
 609         ntb_qp_link_cleanup(qp);
 610
 611         if (nt->transport_link == NTB_LINK_UP)
 612                 schedule_delayed_work(&qp->link_work,
 613                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 614 }
 615
 616 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 617 {
 618         schedule_work(&qp->link_cleanup);
 619 }
 620
 621 static void ntb_transport_link_cleanup(struct ntb_transport *nt)
 622 {
 623         int i;
 624
 625         /* Pass along the info to any clients */
 626         for (i = 0; i < nt->max_qps; i++)
 627                 if (!test_bit(i, &nt->qp_bitmap))
 628                         ntb_qp_link_cleanup(&nt->qps[i]);
 629
 630         if (nt->transport_link == NTB_LINK_DOWN)
 631                 cancel_delayed_work_sync(&nt->link_work);
 632         else
 633                 nt->transport_link = NTB_LINK_DOWN;
 634
 635         /* The scratchpad registers keep the values if the remote side
 636          * goes down, blast them now to give them a sane value the next
 637          * time they are accessed
 638          */
 639         for (i = 0; i < MAX_SPAD; i++)
 640                 ntb_write_local_spad(nt->ndev, i, 0);
 641 }
 642
 643 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 644 {
 645         struct ntb_transport *nt = container_of(work, struct ntb_transport,
 646                                                 link_cleanup);
 647
 648         ntb_transport_link_cleanup(nt);
 649 }
 650
 651 static void ntb_transport_event_callback(void *data, enum ntb_hw_event event)
 652 {
 653         struct ntb_transport *nt = data;
 654
 655         switch (event) {
 656         case NTB_EVENT_HW_LINK_UP:
 657                 schedule_delayed_work(&nt->link_work, 0);
 658                 break;
 659         case NTB_EVENT_HW_LINK_DOWN:
 660                 schedule_work(&nt->link_cleanup);
 661                 break;
 662         default:
 663                 BUG();
 664         }
 665 }
 666
 667 static void ntb_transport_link_work(struct work_struct *work)
 668 {
 669         struct ntb_transport *nt = container_of(work, struct ntb_transport,
 670                                                 link_work.work);
 671         struct ntb_device *ndev = nt->ndev;
 672         struct pci_dev *pdev = ntb_query_pdev(ndev);
 673         u32 val;
 674         int rc, i;
 675
 676         /* send the local info, in the opposite order of the way we read it */
 677         for (i = 0; i < ntb_max_mw(ndev); i++) {
 678                 rc = ntb_write_remote_spad(ndev, MW0_SZ_HIGH + (i * 2),
 679                                            ntb_get_mw_size(ndev, i) >> 32);
 680                 if (rc) {
 681                         dev_err(&pdev->dev, "Error writing %u to remote spad %d\n",
 682                                 (u32)(ntb_get_mw_size(ndev, i) >> 32),
 683                                 MW0_SZ_HIGH + (i * 2));
 684                         goto out;
 685                 }
 686
 687                 rc = ntb_write_remote_spad(ndev, MW0_SZ_LOW + (i * 2),
 688                                            (u32) ntb_get_mw_size(ndev, i));
 689                 if (rc) {
 690                         dev_err(&pdev->dev, "Error writing %u to remote spad %d\n",
 691                                 (u32) ntb_get_mw_size(ndev, i),
 692                                 MW0_SZ_LOW + (i * 2));
 693                         goto out;
 694                 }
 695         }
 696
 697         rc = ntb_write_remote_spad(ndev, NUM_MWS, ntb_max_mw(ndev));
 698         if (rc) {
 699                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 700                         ntb_max_mw(ndev), NUM_MWS);
 701                 goto out;
 702         }
 703
 704         rc = ntb_write_remote_spad(ndev, NUM_QPS, nt->max_qps);
 705         if (rc) {
 706                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 707                         nt->max_qps, NUM_QPS);
 708                 goto out;
 709         }
 710
 711         rc = ntb_write_remote_spad(ndev, VERSION, NTB_TRANSPORT_VERSION);
 712         if (rc) {
 713                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 714                         NTB_TRANSPORT_VERSION, VERSION);
 715                 goto out;
 716         }
 717
 718         /* Query the remote side for its info */
 719         rc = ntb_read_remote_spad(ndev, VERSION, &val);
 720         if (rc) {
 721                 dev_err(&pdev->dev, "Error reading remote spad %d\n", VERSION);
 722                 goto out;
 723         }
 724
 725         if (val != NTB_TRANSPORT_VERSION)
 726                 goto out;
 727         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 728
 729         rc = ntb_read_remote_spad(ndev, NUM_QPS, &val);
 730         if (rc) {
 731                 dev_err(&pdev->dev, "Error reading remote spad %d\n", NUM_QPS);
 732                 goto out;
 733         }
 734
 735         if (val != nt->max_qps)
 736                 goto out;
 737         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 738
 739         rc = ntb_read_remote_spad(ndev, NUM_MWS, &val);
 740         if (rc) {
 741                 dev_err(&pdev->dev, "Error reading remote spad %d\n", NUM_MWS);
 742                 goto out;
 743         }
 744
 745         if (val != ntb_max_mw(ndev))
 746                 goto out;
 747         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 748
 749         for (i = 0; i < ntb_max_mw(ndev); i++) {
 750                 u64 val64;
 751
 752                 rc = ntb_read_remote_spad(ndev, MW0_SZ_HIGH + (i * 2), &val);
 753                 if (rc) {
 754                         dev_err(&pdev->dev, "Error reading remote spad %d\n",
 755                                 MW0_SZ_HIGH + (i * 2));
 756                         goto out1;
 757                 }
 758
 759                 val64 = (u64) val << 32;
 760
 761                 rc = ntb_read_remote_spad(ndev, MW0_SZ_LOW + (i * 2), &val);
 762                 if (rc) {
 763                         dev_err(&pdev->dev, "Error reading remote spad %d\n",
 764                                 MW0_SZ_LOW + (i * 2));
 765                         goto out1;
 766                 }
 767
 768                 val64 |= val;
 769
 770                 dev_dbg(&pdev->dev, "Remote MW%d size = %llu\n", i, val64);
 771
 772                 rc = ntb_set_mw(nt, i, val64);
 773                 if (rc)
 774                         goto out1;
 775         }
 776
 777         nt->transport_link = NTB_LINK_UP;
 778
 779         for (i = 0; i < nt->max_qps; i++) {
 780                 struct ntb_transport_qp *qp = &nt->qps[i];
 781
 782                 ntb_transport_setup_qp_mw(nt, i);
 783
 784                 if (qp->client_ready == NTB_LINK_UP)
 785                         schedule_delayed_work(&qp->link_work, 0);
 786         }
 787
 788         return;
 789
 790 out1:
 791         for (i = 0; i < ntb_max_mw(ndev); i++)
 792                 ntb_free_mw(nt, i);
 793 out:
 794         if (ntb_hw_link_status(ndev))
 795                 schedule_delayed_work(&nt->link_work,
 796                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 797 }
 798
 799 static void ntb_qp_link_work(struct work_struct *work)
 800 {
 801         struct ntb_transport_qp *qp = container_of(work,
 802                                                    struct ntb_transport_qp,
 803                                                    link_work.work);
 804         struct pci_dev *pdev = ntb_query_pdev(qp->ndev);
 805         struct ntb_transport *nt = qp->transport;
 806         int rc, val;
 807
 808         WARN_ON(nt->transport_link != NTB_LINK_UP);
 809
 810         rc = ntb_read_local_spad(nt->ndev, QP_LINKS, &val);
 811         if (rc) {
 812                 dev_err(&pdev->dev, "Error reading spad %d\n", QP_LINKS);
 813                 return;
 814         }
 815
 816         rc = ntb_write_remote_spad(nt->ndev, QP_LINKS, val | 1 << qp->qp_num);
 817         if (rc)
 818                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 819                         val | 1 << qp->qp_num, QP_LINKS);
 820
 821         /* query remote spad for qp ready bits */
 822         rc = ntb_read_remote_spad(nt->ndev, QP_LINKS, &val);
 823         if (rc)
 824                 dev_err(&pdev->dev, "Error reading remote spad %d\n", QP_LINKS);
 825
 826         dev_dbg(&pdev->dev, "Remote QP link status = %x\n", val);
 827
 828         /* See if the remote side is up */
 829         if (1 << qp->qp_num & val) {
 830                 qp->qp_link = NTB_LINK_UP;
 831
 832                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 833                 if (qp->event_handler)
 834                         qp->event_handler(qp->cb_data, NTB_LINK_UP);
 835         } else if (nt->transport_link == NTB_LINK_UP)
 836                 schedule_delayed_work(&qp->link_work,
 837                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 838 }
 839
 840 static int ntb_transport_init_queue(struct ntb_transport *nt,
 841                                     unsigned int qp_num)
 842 {
 843         struct ntb_transport_qp *qp;
 844         unsigned int num_qps_mw, tx_size;
 845         u8 mw_num, mw_max;
 846         u64 qp_offset;
 847
 848         mw_max = ntb_max_mw(nt->ndev);
 849         mw_num = QP_TO_MW(nt->ndev, qp_num);
 850
 851         qp = &nt->qps[qp_num];
 852         qp->qp_num = qp_num;
 853         qp->transport = nt;
 854         qp->ndev = nt->ndev;
 855         qp->qp_link = NTB_LINK_DOWN;
 856         qp->client_ready = NTB_LINK_DOWN;
 857         qp->event_handler = NULL;
 858
 859         if (nt->max_qps % mw_max && mw_num < nt->max_qps % mw_max)
 860                 num_qps_mw = nt->max_qps / mw_max + 1;
 861         else
 862                 num_qps_mw = nt->max_qps / mw_max;
 863
 864         tx_size = (unsigned int) ntb_get_mw_size(qp->ndev, mw_num) / num_qps_mw;
 865         qp_offset = qp_num / mw_max * tx_size;
 866         qp->tx_mw = ntb_get_mw_vbase(nt->ndev, mw_num) + qp_offset;
 867         if (!qp->tx_mw)
 868                 return -EINVAL;
 869
 870         qp->tx_mw_phys = ntb_get_mw_base(qp->ndev, mw_num) + qp_offset;
 871         if (!qp->tx_mw_phys)
 872                 return -EINVAL;
 873
 874         tx_size -= sizeof(struct ntb_rx_info);
 875         qp->rx_info = qp->tx_mw + tx_size;
 876
 877         /* Due to housekeeping, there must be atleast 2 buffs */
 878         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
 879         qp->tx_max_entry = tx_size / qp->tx_max_frame;
 880
 881         if (ntb_query_debugfs(nt->ndev)) {
 882                 char debugfs_name[4];
 883
 884                 snprintf(debugfs_name, 4, "qp%d", qp_num);
 885                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
 886                                                  ntb_query_debugfs(nt->ndev));
 887
 888                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
 889                                                         qp->debugfs_dir, qp,
 890                                                         &ntb_qp_debugfs_stats);
 891         }
 892
 893         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
 894         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
 895
 896         spin_lock_init(&qp->ntb_rx_pend_q_lock);
 897         spin_lock_init(&qp->ntb_rx_free_q_lock);
 898         spin_lock_init(&qp->ntb_tx_free_q_lock);
 899
 900         INIT_LIST_HEAD(&qp->rx_pend_q);
 901         INIT_LIST_HEAD(&qp->rx_free_q);
 902         INIT_LIST_HEAD(&qp->tx_free_q);
 903
 904         return 0;
 905 }
 906
 907 int ntb_transport_init(struct pci_dev *pdev)
 908 {
 909         struct ntb_transport *nt;
 910         int rc, i;
 911
 912         nt = kzalloc(sizeof(struct ntb_transport), GFP_KERNEL);
 913         if (!nt)
 914                 return -ENOMEM;
 915
 916         nt->ndev = ntb_register_transport(pdev, nt);
 917         if (!nt->ndev) {
 918                 rc = -EIO;
 919                 goto err;
 920         }
 921
 922         nt->mw = kcalloc(ntb_max_mw(nt->ndev), sizeof(struct ntb_transport_mw),
 923                          GFP_KERNEL);
 924         if (!nt->mw) {
 925                 rc = -ENOMEM;
 926                 goto err1;
 927         }
 928
 929         if (max_num_clients)
 930                 nt->max_qps = min(ntb_max_cbs(nt->ndev), max_num_clients);
 931         else
 932                 nt->max_qps = min(ntb_max_cbs(nt->ndev), ntb_max_mw(nt->ndev));
 933
 934         nt->qps = kcalloc(nt->max_qps, sizeof(struct ntb_transport_qp),
 935                           GFP_KERNEL);
 936         if (!nt->qps) {
 937                 rc = -ENOMEM;
 938                 goto err2;
 939         }
 940
 941         nt->qp_bitmap = ((u64) 1 << nt->max_qps) - 1;
 942
 943         for (i = 0; i < nt->max_qps; i++) {
 944                 rc = ntb_transport_init_queue(nt, i);
 945                 if (rc)
 946                         goto err3;
 947         }
 948
 949         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
 950         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
 951
 952         rc = ntb_register_event_callback(nt->ndev,
 953                                          ntb_transport_event_callback);
 954         if (rc)
 955                 goto err3;
 956
 957         INIT_LIST_HEAD(&nt->client_devs);
 958         rc = ntb_bus_init(nt);
 959         if (rc)
 960                 goto err4;
 961
 962         if (ntb_hw_link_status(nt->ndev))
 963                 schedule_delayed_work(&nt->link_work, 0);
 964
 965         return 0;
 966
 967 err4:
 968         ntb_unregister_event_callback(nt->ndev);
 969 err3:
 970         kfree(nt->qps);
 971 err2:
 972         kfree(nt->mw);
 973 err1:
 974         ntb_unregister_transport(nt->ndev);
 975 err:
 976         kfree(nt);
 977         return rc;
 978 }
 979
 980 void ntb_transport_free(void *transport)
 981 {
 982         struct ntb_transport *nt = transport;
 983         struct ntb_device *ndev = nt->ndev;
 984         int i;
 985
 986         ntb_transport_link_cleanup(nt);
 987
 988         /* verify that all the qp's are freed */
 989         for (i = 0; i < nt->max_qps; i++) {
 990                 if (!test_bit(i, &nt->qp_bitmap))
 991                         ntb_transport_free_queue(&nt->qps[i]);
 992                 debugfs_remove_recursive(nt->qps[i].debugfs_dir);
 993         }
 994
 995         ntb_bus_remove(nt);
 996
 997         cancel_delayed_work_sync(&nt->link_work);
 998
 999         ntb_unregister_event_callback(ndev);
1000
1001         for (i = 0; i < ntb_max_mw(ndev); i++)
1002                 ntb_free_mw(nt, i);
1003
1004         kfree(nt->qps);
1005         kfree(nt->mw);
1006         ntb_unregister_transport(ndev);
1007         kfree(nt);
1008 }
1009
1010 static void ntb_rx_copy_callback(void *data)
1011 {
1012         struct ntb_queue_entry *entry = data;
1013         struct ntb_transport_qp *qp = entry->qp;
1014         void *cb_data = entry->cb_data;
1015         unsigned int len = entry->len;
1016         struct ntb_payload_header *hdr = entry->rx_hdr;
1017
1018         /* Ensure that the data is fully copied out before clearing the flag */
1019         wmb();
1020         hdr->flags = 0;
1021
1022         iowrite32(entry->index, &qp->rx_info->entry);
1023
1024         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1025
1026         if (qp->rx_handler && qp->client_ready == NTB_LINK_UP)
1027                 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1028 }
1029
1030 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1031 {
1032         void *buf = entry->buf;
1033         size_t len = entry->len;
1034
1035         memcpy(buf, offset, len);
1036
1037         ntb_rx_copy_callback(entry);
1038 }
1039
1040 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset,
1041                          size_t len)
1042 {
1043         struct dma_async_tx_descriptor *txd;
1044         struct ntb_transport_qp *qp = entry->qp;
1045         struct dma_chan *chan = qp->dma_chan;
1046         struct dma_device *device;
1047         size_t pay_off, buff_off;
1048         struct dmaengine_unmap_data *unmap;
1049         dma_cookie_t cookie;
1050         void *buf = entry->buf;
1051
1052         entry->len = len;
1053
1054         if (!chan)
1055                 goto err;
1056
1057         if (len < copy_bytes)
1058                 goto err_wait;
1059
1060         device = chan->device;
1061         pay_off = (size_t) offset & ~PAGE_MASK;
1062         buff_off = (size_t) buf & ~PAGE_MASK;
1063
1064         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1065                 goto err_wait;
1066
1067         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1068         if (!unmap)
1069                 goto err_wait;
1070
1071         unmap->len = len;
1072         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1073                                       pay_off, len, DMA_TO_DEVICE);
1074         if (dma_mapping_error(device->dev, unmap->addr[0]))
1075                 goto err_get_unmap;
1076
1077         unmap->to_cnt = 1;
1078
1079         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1080                                       buff_off, len, DMA_FROM_DEVICE);
1081         if (dma_mapping_error(device->dev, unmap->addr[1]))
1082                 goto err_get_unmap;
1083
1084         unmap->from_cnt = 1;
1085
1086         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1087                                              unmap->addr[0], len,
1088                                              DMA_PREP_INTERRUPT);
1089         if (!txd)
1090                 goto err_get_unmap;
1091
1092         txd->callback = ntb_rx_copy_callback;
1093         txd->callback_param = entry;
1094         dma_set_unmap(txd, unmap);
1095
1096         cookie = dmaengine_submit(txd);
1097         if (dma_submit_error(cookie))
1098                 goto err_set_unmap;
1099
1100         dmaengine_unmap_put(unmap);
1101
1102         qp->last_cookie = cookie;
1103
1104         qp->rx_async++;
1105
1106         return;
1107
1108 err_set_unmap:
1109         dmaengine_unmap_put(unmap);
1110 err_get_unmap:
1111         dmaengine_unmap_put(unmap);
1112 err_wait:
1113         /* If the callbacks come out of order, the writing of the index to the
1114          * last completed will be out of order.  This may result in the
1115          * receive stalling forever.
1116          */
1117         dma_sync_wait(chan, qp->last_cookie);
1118 err:
1119         ntb_memcpy_rx(entry, offset);
1120         qp->rx_memcpy++;
1121 }
1122
1123 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1124 {
1125         struct ntb_payload_header *hdr;
1126         struct ntb_queue_entry *entry;
1127         void *offset;
1128
1129         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1130         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1131
1132         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1133         if (!entry) {
1134                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1135                         "no buffer - HDR ver %u, len %d, flags %x\n",
1136                         hdr->ver, hdr->len, hdr->flags);
1137                 qp->rx_err_no_buf++;
1138                 return -ENOMEM;
1139         }
1140
1141         if (!(hdr->flags & DESC_DONE_FLAG)) {
1142                 ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1143                              &qp->rx_pend_q);
1144                 qp->rx_ring_empty++;
1145                 return -EAGAIN;
1146         }
1147
1148         if (hdr->ver != (u32) qp->rx_pkts) {
1149                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1150                         "qp %d: version mismatch, expected %llu - got %u\n",
1151                         qp->qp_num, qp->rx_pkts, hdr->ver);
1152                 ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1153                              &qp->rx_pend_q);
1154                 qp->rx_err_ver++;
1155                 return -EIO;
1156         }
1157
1158         if (hdr->flags & LINK_DOWN_FLAG) {
1159                 ntb_qp_link_down(qp);
1160
1161                 goto err;
1162         }
1163
1164         dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1165                 "rx offset %u, ver %u - %d payload received, buf size %d\n",
1166                 qp->rx_index, hdr->ver, hdr->len, entry->len);
1167
1168         qp->rx_bytes += hdr->len;
1169         qp->rx_pkts++;
1170
1171         if (hdr->len > entry->len) {
1172                 qp->rx_err_oflow++;
1173                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1174                         "RX overflow! Wanted %d got %d\n",
1175                         hdr->len, entry->len);
1176
1177                 goto err;
1178         }
1179
1180         entry->index = qp->rx_index;
1181         entry->rx_hdr = hdr;
1182
1183         ntb_async_rx(entry, offset, hdr->len);
1184
1185 out:
1186         qp->rx_index++;
1187         qp->rx_index %= qp->rx_max_entry;
1188
1189         return 0;
1190
1191 err:
1192         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry, &qp->rx_pend_q);
1193         /* Ensure that the data is fully copied out before clearing the flag */
1194         wmb();
1195         hdr->flags = 0;
1196         iowrite32(qp->rx_index, &qp->rx_info->entry);
1197
1198         goto out;
1199 }
1200
1201 static int ntb_transport_rxc_db(void *data, int db_num)
1202 {
1203         struct ntb_transport_qp *qp = data;
1204         int rc, i;
1205
1206         dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%s: doorbell %d received\n",
1207                 __func__, db_num);
1208
1209         /* Limit the number of packets processed in a single interrupt to
1210          * provide fairness to others
1211          */
1212         for (i = 0; i < qp->rx_max_entry; i++) {
1213                 rc = ntb_process_rxc(qp);
1214                 if (rc)
1215                         break;
1216         }
1217
1218         if (qp->dma_chan)
1219                 dma_async_issue_pending(qp->dma_chan);
1220
1221         return i;
1222 }
1223
1224 static void ntb_tx_copy_callback(void *data)
1225 {
1226         struct ntb_queue_entry *entry = data;
1227         struct ntb_transport_qp *qp = entry->qp;
1228         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1229
1230         /* Ensure that the data is fully copied out before setting the flags */
1231         wmb();
1232         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1233
1234         ntb_ring_doorbell(qp->ndev, qp->qp_num);
1235
1236         /* The entry length can only be zero if the packet is intended to be a
1237          * "link down" or similar.  Since no payload is being sent in these
1238          * cases, there is nothing to add to the completion queue.
1239          */
1240         if (entry->len > 0) {
1241                 qp->tx_bytes += entry->len;
1242
1243                 if (qp->tx_handler)
1244                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1245                                        entry->len);
1246         }
1247
1248         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1249 }
1250
1251 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1252 {
1253         memcpy_toio(offset, entry->buf, entry->len);
1254
1255         ntb_tx_copy_callback(entry);
1256 }
1257
1258 static void ntb_async_tx(struct ntb_transport_qp *qp,
1259                          struct ntb_queue_entry *entry)
1260 {
1261         struct ntb_payload_header __iomem *hdr;
1262         struct dma_async_tx_descriptor *txd;
1263         struct dma_chan *chan = qp->dma_chan;
1264         struct dma_device *device;
1265         size_t dest_off, buff_off;
1266         struct dmaengine_unmap_data *unmap;
1267         dma_addr_t dest;
1268         dma_cookie_t cookie;
1269         void __iomem *offset;
1270         size_t len = entry->len;
1271         void *buf = entry->buf;
1272
1273         offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1274         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1275         entry->tx_hdr = hdr;
1276
1277         iowrite32(entry->len, &hdr->len);
1278         iowrite32((u32) qp->tx_pkts, &hdr->ver);
1279
1280         if (!chan)
1281                 goto err;
1282
1283         if (len < copy_bytes)
1284                 goto err;
1285
1286         device = chan->device;
1287         dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1288         buff_off = (size_t) buf & ~PAGE_MASK;
1289         dest_off = (size_t) dest & ~PAGE_MASK;
1290
1291         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1292                 goto err;
1293
1294         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1295         if (!unmap)
1296                 goto err;
1297
1298         unmap->len = len;
1299         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1300                                       buff_off, len, DMA_TO_DEVICE);
1301         if (dma_mapping_error(device->dev, unmap->addr[0]))
1302                 goto err_get_unmap;
1303
1304         unmap->to_cnt = 1;
1305
1306         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1307                                              DMA_PREP_INTERRUPT);
1308         if (!txd)
1309                 goto err_get_unmap;
1310
1311         txd->callback = ntb_tx_copy_callback;
1312         txd->callback_param = entry;
1313         dma_set_unmap(txd, unmap);
1314
1315         cookie = dmaengine_submit(txd);
1316         if (dma_submit_error(cookie))
1317                 goto err_set_unmap;
1318
1319         dmaengine_unmap_put(unmap);
1320
1321         dma_async_issue_pending(chan);
1322         qp->tx_async++;
1323
1324         return;
1325 err_set_unmap:
1326         dmaengine_unmap_put(unmap);
1327 err_get_unmap:
1328         dmaengine_unmap_put(unmap);
1329 err:
1330         ntb_memcpy_tx(entry, offset);
1331         qp->tx_memcpy++;
1332 }
1333
1334 static int ntb_process_tx(struct ntb_transport_qp *qp,
1335                           struct ntb_queue_entry *entry)
1336 {
1337         dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%lld - tx %u, entry len %d flags %x buff %p\n",
1338                 qp->tx_pkts, qp->tx_index, entry->len, entry->flags,
1339                 entry->buf);
1340         if (qp->tx_index == qp->remote_rx_info->entry) {
1341                 qp->tx_ring_full++;
1342                 return -EAGAIN;
1343         }
1344
1345         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1346                 if (qp->tx_handler)
1347                         qp->tx_handler(qp->cb_data, qp, NULL, -EIO);
1348
1349                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1350                              &qp->tx_free_q);
1351                 return 0;
1352         }
1353
1354         ntb_async_tx(qp, entry);
1355
1356         qp->tx_index++;
1357         qp->tx_index %= qp->tx_max_entry;
1358
1359         qp->tx_pkts++;
1360
1361         return 0;
1362 }
1363
1364 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1365 {
1366         struct pci_dev *pdev = ntb_query_pdev(qp->ndev);
1367         struct ntb_queue_entry *entry;
1368         int i, rc;
1369
1370         if (qp->qp_link == NTB_LINK_DOWN)
1371                 return;
1372
1373         qp->qp_link = NTB_LINK_DOWN;
1374         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
1375
1376         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1377                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1378                 if (entry)
1379                         break;
1380                 msleep(100);
1381         }
1382
1383         if (!entry)
1384                 return;
1385
1386         entry->cb_data = NULL;
1387         entry->buf = NULL;
1388         entry->len = 0;
1389         entry->flags = LINK_DOWN_FLAG;
1390
1391         rc = ntb_process_tx(qp, entry);
1392         if (rc)
1393                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1394                         qp->qp_num);
1395 }
1396
1397 /**
1398  * ntb_transport_create_queue - Create a new NTB transport layer queue
1399  * @rx_handler: receive callback function
1400  * @tx_handler: transmit callback function
1401  * @event_handler: event callback function
1402  *
1403  * Create a new NTB transport layer queue and provide the queue with a callback
1404  * routine for both transmit and receive.  The receive callback routine will be
1405  * used to pass up data when the transport has received it on the queue.   The
1406  * transmit callback routine will be called when the transport has completed the
1407  * transmission of the data on the queue and the data is ready to be freed.
1408  *
1409  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1410  */
1411 struct ntb_transport_qp *
1412 ntb_transport_create_queue(void *data, struct pci_dev *pdev,
1413                            const struct ntb_queue_handlers *handlers)
1414 {
1415         struct ntb_queue_entry *entry;
1416         struct ntb_transport_qp *qp;
1417         struct ntb_transport *nt;
1418         unsigned int free_queue;
1419         int rc, i;
1420
1421         nt = ntb_find_transport(pdev);
1422         if (!nt)
1423                 goto err;
1424
1425         free_queue = ffs(nt->qp_bitmap);
1426         if (!free_queue)
1427                 goto err;
1428
1429         /* decrement free_queue to make it zero based */
1430         free_queue--;
1431
1432         clear_bit(free_queue, &nt->qp_bitmap);
1433
1434         qp = &nt->qps[free_queue];
1435         qp->cb_data = data;
1436         qp->rx_handler = handlers->rx_handler;
1437         qp->tx_handler = handlers->tx_handler;
1438         qp->event_handler = handlers->event_handler;
1439
1440         dmaengine_get();
1441         qp->dma_chan = dma_find_channel(DMA_MEMCPY);
1442         if (!qp->dma_chan) {
1443                 dmaengine_put();
1444                 dev_info(&pdev->dev, "Unable to allocate DMA channel, using CPU instead\n");
1445         }
1446
1447         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1448                 entry = kzalloc(sizeof(struct ntb_queue_entry), GFP_ATOMIC);
1449                 if (!entry)
1450                         goto err1;
1451
1452                 entry->qp = qp;
1453                 ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry,
1454                              &qp->rx_free_q);
1455         }
1456
1457         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1458                 entry = kzalloc(sizeof(struct ntb_queue_entry), GFP_ATOMIC);
1459                 if (!entry)
1460                         goto err2;
1461
1462                 entry->qp = qp;
1463                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1464                              &qp->tx_free_q);
1465         }
1466
1467         rc = ntb_register_db_callback(qp->ndev, free_queue, qp,
1468                                       ntb_transport_rxc_db);
1469         if (rc)
1470                 goto err2;
1471
1472         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1473
1474         return qp;
1475
1476 err2:
1477         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1478                 kfree(entry);
1479 err1:
1480         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1481                 kfree(entry);
1482         if (qp->dma_chan)
1483                 dmaengine_put();
1484         set_bit(free_queue, &nt->qp_bitmap);
1485 err:
1486         return NULL;
1487 }
1488 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1489
1490 /**
1491  * ntb_transport_free_queue - Frees NTB transport queue
1492  * @qp: NTB queue to be freed
1493  *
1494  * Frees NTB transport queue
1495  */
1496 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1497 {
1498         struct pci_dev *pdev;
1499         struct ntb_queue_entry *entry;
1500
1501         if (!qp)
1502                 return;
1503
1504         pdev = ntb_query_pdev(qp->ndev);
1505
1506         if (qp->dma_chan) {
1507                 struct dma_chan *chan = qp->dma_chan;
1508                 /* Putting the dma_chan to NULL will force any new traffic to be
1509                  * processed by the CPU instead of the DAM engine
1510                  */
1511                 qp->dma_chan = NULL;
1512
1513                 /* Try to be nice and wait for any queued DMA engine
1514                  * transactions to process before smashing it with a rock
1515                  */
1516                 dma_sync_wait(chan, qp->last_cookie);
1517                 dmaengine_terminate_all(chan);
1518                 dmaengine_put();
1519         }
1520
1521         ntb_unregister_db_callback(qp->ndev, qp->qp_num);
1522
1523         cancel_delayed_work_sync(&qp->link_work);
1524
1525         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1526                 kfree(entry);
1527
1528         while ((entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q))) {
1529                 dev_warn(&pdev->dev, "Freeing item from a non-empty queue\n");
1530                 kfree(entry);
1531         }
1532
1533         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1534                 kfree(entry);
1535
1536         set_bit(qp->qp_num, &qp->transport->qp_bitmap);
1537
1538         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1539 }
1540 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1541
1542 /**
1543  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1544  * @qp: NTB queue to be freed
1545  * @len: pointer to variable to write enqueued buffers length
1546  *
1547  * Dequeues unused buffers from receive queue.  Should only be used during
1548  * shutdown of qp.
1549  *
1550  * RETURNS: NULL error value on error, or void* for success.
1551  */
1552 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1553 {
1554         struct ntb_queue_entry *entry;
1555         void *buf;
1556
1557         if (!qp || qp->client_ready == NTB_LINK_UP)
1558                 return NULL;
1559
1560         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1561         if (!entry)
1562                 return NULL;
1563
1564         buf = entry->cb_data;
1565         *len = entry->len;
1566
1567         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1568
1569         return buf;
1570 }
1571 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1572
1573 /**
1574  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1575  * @qp: NTB transport layer queue the entry is to be enqueued on
1576  * @cb: per buffer pointer for callback function to use
1577  * @data: pointer to data buffer that incoming packets will be copied into
1578  * @len: length of the data buffer
1579  *
1580  * Enqueue a new receive buffer onto the transport queue into which a NTB
1581  * payload can be received into.
1582  *
1583  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1584  */
1585 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1586                              unsigned int len)
1587 {
1588         struct ntb_queue_entry *entry;
1589
1590         if (!qp)
1591                 return -EINVAL;
1592
1593         entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q);
1594         if (!entry)
1595                 return -ENOMEM;
1596
1597         entry->cb_data = cb;
1598         entry->buf = data;
1599         entry->len = len;
1600
1601         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry, &qp->rx_pend_q);
1602
1603         return 0;
1604 }
1605 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1606
1607 /**
1608  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1609  * @qp: NTB transport layer queue the entry is to be enqueued on
1610  * @cb: per buffer pointer for callback function to use
1611  * @data: pointer to data buffer that will be sent
1612  * @len: length of the data buffer
1613  *
1614  * Enqueue a new transmit buffer onto the transport queue from which a NTB
1615  * payload will be transmitted.  This assumes that a lock is being held to
1616  * serialize access to the qp.
1617  *
1618  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1619  */
1620 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1621                              unsigned int len)
1622 {
1623         struct ntb_queue_entry *entry;
1624         int rc;
1625
1626         if (!qp || qp->qp_link != NTB_LINK_UP || !len)
1627                 return -EINVAL;
1628
1629         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1630         if (!entry) {
1631                 qp->tx_err_no_buf++;
1632                 return -ENOMEM;
1633         }
1634
1635         entry->cb_data = cb;
1636         entry->buf = data;
1637         entry->len = len;
1638         entry->flags = 0;
1639
1640         rc = ntb_process_tx(qp, entry);
1641         if (rc)
1642                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1643                              &qp->tx_free_q);
1644
1645         return rc;
1646 }
1647 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1648
1649 /**
1650  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1651  * @qp: NTB transport layer queue to be enabled
1652  *
1653  * Notify NTB transport layer of client readiness to use queue
1654  */
1655 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1656 {
1657         if (!qp)
1658                 return;
1659
1660         qp->client_ready = NTB_LINK_UP;
1661
1662         if (qp->transport->transport_link == NTB_LINK_UP)
1663                 schedule_delayed_work(&qp->link_work, 0);
1664 }
1665 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
1666
1667 /**
1668  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1669  * @qp: NTB transport layer queue to be disabled
1670  *
1671  * Notify NTB transport layer of client's desire to no longer receive data on
1672  * transport queue specified.  It is the client's responsibility to ensure all
1673  * entries on queue are purged or otherwise handled appropriately.
1674  */
1675 void ntb_transport_link_down(struct ntb_transport_qp *qp)
1676 {
1677         struct pci_dev *pdev;
1678         int rc, val;
1679
1680         if (!qp)
1681                 return;
1682
1683         pdev = ntb_query_pdev(qp->ndev);
1684         qp->client_ready = NTB_LINK_DOWN;
1685
1686         rc = ntb_read_local_spad(qp->ndev, QP_LINKS, &val);
1687         if (rc) {
1688                 dev_err(&pdev->dev, "Error reading spad %d\n", QP_LINKS);
1689                 return;
1690         }
1691
1692         rc = ntb_write_remote_spad(qp->ndev, QP_LINKS,
1693                                    val & ~(1 << qp->qp_num));
1694         if (rc)
1695                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
1696                         val & ~(1 << qp->qp_num), QP_LINKS);
1697
1698         if (qp->qp_link == NTB_LINK_UP)
1699                 ntb_send_link_down(qp);
1700         else
1701                 cancel_delayed_work_sync(&qp->link_work);
1702 }
1703 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
1704
1705 /**
1706  * ntb_transport_link_query - Query transport link state
1707  * @qp: NTB transport layer queue to be queried
1708  *
1709  * Query connectivity to the remote system of the NTB transport queue
1710  *
1711  * RETURNS: true for link up or false for link down
1712  */
1713 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
1714 {
1715         if (!qp)
1716                 return false;
1717
1718         return qp->qp_link == NTB_LINK_UP;
1719 }
1720 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
1721
1722 /**
1723  * ntb_transport_qp_num - Query the qp number
1724  * @qp: NTB transport layer queue to be queried
1725  *
1726  * Query qp number of the NTB transport queue
1727  *
1728  * RETURNS: a zero based number specifying the qp number
1729  */
1730 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1731 {
1732         if (!qp)
1733                 return 0;
1734
1735         return qp->qp_num;
1736 }
1737 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
1738
1739 /**
1740  * ntb_transport_max_size - Query the max payload size of a qp
1741  * @qp: NTB transport layer queue to be queried
1742  *
1743  * Query the maximum payload size permissible on the given qp
1744  *
1745  * RETURNS: the max payload size of a qp
1746  */
1747 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
1748 {
1749         unsigned int max;
1750
1751         if (!qp)
1752                 return 0;
1753
1754         if (!qp->dma_chan)
1755                 return qp->tx_max_frame - sizeof(struct ntb_payload_header);
1756
1757         /* If DMA engine usage is possible, try to find the max size for that */
1758         max = qp->tx_max_frame - sizeof(struct ntb_payload_header);
1759         max -= max % (1 << qp->dma_chan->device->copy_align);
1760
1761         return max;
1762 }
1763 EXPORT_SYMBOL_GPL(ntb_transport_max_size);